About me
I am currently a postdoctoral researcher at LISN (Université Paris-Saclay), working on the evaluation of LLMs for clinical applications in the PARTAGES project.
My research lies at the intersection of Natural Language Processing and health, with a focus on the extraction and structuring of clinical information from unstructured text, and on ensuring the reliability and trustworthiness of AI systems in real-world clinical applications.
I am particularly interested in the evaluation of LLMs in clinical contexts, the design of robust benchmarks, and the development of methods that combine linguistic knowledge, statistical learning, and domain expertise.
Research Interests
Evaluation of LLMs for Clinical Text Processing
Understanding how and when LLMs can be reliably used for clinical information extraction and generation.
Model Usage
Applicability of LLMs to information extraction and generation tasks, including prompt design and the role of external knowledge and context.
Evaluation
Design of robust evaluation frameworks combining quantitative metrics and qualitative analysis, including LLM-as-a-judge approaches.
Reliability
Calibration, uncertainty estimation, and bias analysis in clinical settings.
Data
Construction of reliable ground truth datasets in complex and costly clinical annotation environments.
Description
This research investigates the use of LLMs for clinical text processing, with a particular focus on their applicability to information extraction and generation tasks. Key questions include how to effectively prompt LLMs for clinical applications, the role of external knowledge and contextual information, and the extent to which these models can be reliably deployed in practice.
A central aspect of this work is the evaluation of LLM outputs: designing robust evaluation frameworks that combine quantitative metrics and qualitative analysis, exploring the use of LLMs as evaluators (LLM-as-a-judge), and addressing challenges related to calibration, uncertainty estimation, and bias. Particular attention is given to the construction of reliable ground truth data in clinical settings, where annotation is costly and inherently complex.
Information Extraction in the Biomedical Domain
Extracting structured clinical knowledge from heterogeneous biomedical texts.
Tasks
Named entity recognition, relation extraction, and normalization to medical ontologies.
Data Sources
Scientific literature, electronic health records, and medical prescriptions, each with different linguistic challenges.
Models
Transformer-based architectures, LLMs, and knowledge graph integration.
Linguistics
Leveraging lexical, syntactic, and semantic features to improve extraction performance.
Description
This line of research focuses on the extraction of structured information from biomedical and clinical texts, including tasks such as named entity recognition, relation extraction, and normalization to medical ontologies. It considers a variety of data sources, ranging from scientific literature to electronic health records and medical prescriptions, each with distinct linguistic and domain-specific challenges.
The work explores the use of different modeling approaches, including transformer-based architectures, large language models, and knowledge graph integration. It also investigates the contribution of linguistic features (lexical, syntactic, and semantic) to improve extraction performance. A key objective is to better understand what types of information can be reliably extracted, under which conditions, and with what level of accuracy for downstream clinical use.
Professional Experience
Postdoctoral Researcher — LISN, Paris-Saclay
- Participation in the PARTAGES project
- LLM evaluation for different clinical use cases
Data Scientist — Doctolib
- Development and calibration of LLM evaluation scorers for clinical summaries
- Prompt engineering and A/B testing for information extraction & structuration
- Guidelines design for annotation campaigns for clinical data
- Deployment of LLM-based systems on AWS, GCP, ArgoCD
Data Scientist — Posos
- Fine-tuning transformer models for biomedical NER
- Entity linking with medical terminologies (e.g. SNOMED, LOINC)
- Development of production API for automatic medical annotation
- Authored scientific and popular science articles on biomedical NLP
PhD Researcher — LORIA (Université de Lorraine, CNRS, Inria)
- Domain: Biomedical AI, NLP, Information Extraction
- Thesis: Biomedical Event Extraction based on Transformers and Knowledge Graphs
Teaching
Temporary Teaching and Research Associate
Total HEQTD: 165.5h
Télécom Nancy — LORIA (Université de Lorraine, CNRS, Inria)
Publications
2025
-
Confabulations from ACL Publications (CAP): A Dataset for Scientific Hallucination Detection
15th Language Resources and Evaluation Conference (LREC)
-
Automatic Posology Structuration: What role for LLMs?
32nd MLP-LLM / CORIA-TALN
2023
-
How Much do Knowledge Graphs Impact Transformer Models for Extracting Biomedical Events?
61st Annual Meeting of the Association for Computational Linguistics (ACL)
-
Adding Linguistic Information to Transformer Models Improves Biomedical Event Detection?
18th Conference on Computer Science and Intelligence Systems (FedCSIS)
-
Fine-tuning Pre-trained Transformer Language Models for Biomedical Event Trigger Detection
DL4NLP - Extraction et Gestion de Connaissances (EGC)
Scientific Activities
Medical Language Processing in the era of Large Language Models
Hallucinations and Related Observable Overgeneration Mistakes in Crosslingual Analyses of Publication